TabPFN

TabPFN is a foundation model for tabular data from Prior Labs that works like a GPT for spreadsheets: pre-trained once on synthetic tables, it performs classification and regression on new datasets in a single forward pass with no training step. Version 2 was published in Nature in January 2025 (Hollmann et al.) and handles datasets up to around 10,000 rows, making it a genuine alternative to gradient-boosted trees for small-to-medium tabular problems.

License details matter here and are stated honestly: v2 weights are released under a modified Apache license requiring attribution, while the newer 2.5 and 3.x versions carry non-commercial terms. Also note that SAP acquired Prior Labs in May 2026 (reported by third-party coverage), which may shape future licensing. Despite the caveats, TabPFN remains the reference implementation of the tabular-foundation-model idea and is widely used for fast prototyping in data science.